The promise and the trap
Journalism runs on two things AI is unevenly matched to: speed and accuracy. AI is genuinely fast at the speed half — reading thousands of pages, transcribing interviews, suggesting angles — and genuinely dangerous on the accuracy half, because language models fabricate quotes, sources, and statistics with total confidence. The reporter’s job is to harvest the speed without ever trusting the facts. Every workflow below assumes the same iron rule: AI points you at things, you verify them against primary sources, and nothing reaches print unchecked.
Workflows that fit a newsroom
Document analysis for investigations. Faced with a leaked dump of emails, contracts, or filings, load them into a retrieval setup and ask targeted questions: who approved this payment, when did the policy change, which names recur. The model surfaces leads and quotes the source passage; you then read the underlying document yourself before reporting it.
Interview transcript summarisation. Auto-transcribe an interview, then ask for the key admissions, contradictions, and the most quotable lines with timestamps. You jump straight to the moments that matter and confirm each quote against the audio before using it.
Headline and framing testing. Ask for ten headline variations across registers — straight, punchy, SEO-friendly — and a critique of which oversell or mislead. It is an idea generator and a sanity check against clickbait, not the final call.
Background briefing. Before an interview, ask the model to outline the known context and the obvious questions, then go beyond them. Verify any fact it states about the subject; it may be wrong or out of date.
Verification guardrails
Treat every factual output as a hallucination until confirmed against a primary source — especially quotes, numbers, dates, and named sources. Protect confidential material: do not upload sensitive source documents to consumer tools that may retain them; use offline models or manual analysis. Rewrite any AI draft in your own voice and judgement, and disclose AI assistance where your outlet requires it. The byline carries accountability that no model shares, so the standard is unchanged: if you cannot stand behind a fact, it does not run.